PoseGRAF: Geometric-Reinforced Adaptive Fusion...

---

📌 Introduction

PoseGRAF is a novel framework that addresses key challenges in monocular 3D human-pose estimation:

• Dual Graph Convolution – simultaneously processes joint and bone graphs under geometric constraints.
• Geometry-Aware Fusion – explicitly models the relationships between bone directions and joint angles.
• Dynamic Feature Integration – adaptively fuses positional and geometric features via attention.
• Topology-Enhanced Transformer – injects anatomical constraints directly into the transformer architecture.

PoseGRAF achieves state-of-the-art performance on standard benchmarks and generalises well to in-the-wild scenarios.

---

🚀 Key Results

| Metric | Dataset / Protocol | Result | |--------|-------------------|--------| | MPJPE | Human3.6M (CPN detections) | 48.1 mm | | P-MPJPE | Human3.6M | 38.3 mm | | PCK | MPI-INF-3DHP | 87.2 % |

PoseGRAF is robust to occlusion and fast motion.

---

📦 Installation

# Clone the repository
git clone https://github.com/iCityLab/PoseGRAF.git
cd PoseGRAF

# Create the conda environment
conda env create -f environment.yml


---

## 📝 Acknowledgements

This codebase builds on the excellent work below—many thanks to the authors for open-sourcing their projects.

* [DGFormer](https://github.com/czmmmm/DGFormer)
* [PoseFormer](https://github.com/zczcwh/PoseFormer)
* [GraphMLP](https://github.com/Vegetebird/GraphMLP)
* [MeTDDI](https://github.com/LabWeng/MeTDDI)

---

## 🏋️‍♂️ Training from Scratch

Logs, checkpoints and auxiliary files will be written to `./checkpoint`.

#### Human3.6M

```bash
python main.py --train --model PoseGRAF --layers 6 --nepoch 40 --gpu 0

*Note — our training script follows the convention of VideoPose3D. Please refer to their project page for further information.

---

🎬 Pose-Estimation Comparison

<p align="center"> <img src="PoseGRAF/images/dance.png" width="400" alt="2D pose"/> <img src="PoseGRAF/images/dance3D.png" width="400" alt="3D pose"/> </p>

---

🌄 In-the-Wild Demo

1 · Install FFmpeg

Download the latest FFmpeg build from https://ffmpeg.org/download.html and ensure the executable is on your PATH.

2 · Create the Side-by-Side Comparison Video

# ❶ Convert 2D key-point renders to a video
ffmpeg -framerate 30 -start_number 0 -i "%04d_2D.png" -vf "scale=-2:872,fps=100" -c:v libx264 -crf 23 -pix_fmt yuv420p pose2D.mp4

# ❷ Convert 3D key-point renders to a video
ffmpeg -framerate 30 -start_number 0 -i "%04d_3D.png" -vf "scale=-2:872,fps=100" -c:v libx264 -crf 23 -pix_fmt yuv420p pose3D.mp4

# ❸ Concatenate the two videos horizontally
ffmpeg -i pose2D.mp4 -i pose3D.mp4 -filter_complex "hstack=inputs=2" -c:v libx264 -crf 23 -pix_fmt yuv420p pose_comparison.mp4

---

🕺 Animated Dance GIF